Creating Models

There are three ways to create new models and model versions: through the Model Hub web UI, through execution outputs, and by interacting with the Valohai API.

Web application

Creating a Model

On the Model Hub, click on “Create Model” and create a new model.

You only need to choose a name now, but note that this is the only time you can change the technical model:// URI (also known as a ‘slug’) used to refer to this model in execution inputs and other places within your Valohai projects.

Creating a Model Version

The Model Hub enables the creation of model versions as well. Open the overview page for a model, and click on “Create Version”. You can search for and add files from your project data library to the model version.

From execution outputs

Create the model using the Model Hub or API before adding versions to it with executions. Make note of its model://<model-name> URI.

Creating a Model Version

Use the valohai.model-versions metadata statement to create model versions from execution or task outputs. Create a metadata statement for each of the outputs to be added to the new model version. A single execution or task can add at most one new version per model.

import valohai
import json

metadata = {
    "model.pkl": {
        "factory": "eu-02",
        "valohai.tags": ["prod", "lemonade"],
        "valohai.model-versions": ["model://<model-name>/"]
    },
}

# save all the models separately;
save_path = "/valohai/outputs/model.pkl"
model.save(save_path)

metadata_path = "/valohai/outputs/valohai.metadata.jsonl"
with open(metadata_path, "w") as outfile:
    for file_name, file_metadata in metadata.items():
        json.dump({"file": file_name, "metadata": file_metadata}, outfile)
        outfile.write("\n")

You can add tags or release notes to model versions created this way with an extended valohai.model-versions data structure:

metadata = {
    "model.pkl":{
        "valohai.model-versions": [{
            "model_uri": "model://<model-name>/",
            "model_version_tags": ["green", "cucumber"],
            "model_release_note": "100% freshly processed"
        }]
}}

The last output to be processed that has non-empty values for these fields will contribute them to the resulting model version.

Legacy Approach of Creating a Model Version

In the legacy approach, metadata was stored in separate files for each model output. This required creating a dedicated metadata file per file.

For example:

import json
metadata = {
    "valohai.model-versions": ["model://<model-name>/"]
}

save_path = '/valohai/outputs/model.pkl'
model.save(save_path)

metadata_path = '/valohai/outputs/model.pkl.metadata.json'
with open(metadata_path, 'w') as outfile:
    json.dump(metadata, outfile)

This approach still works, so if you have it in your project, you can continue using it. However, the JSONL-based method is recommended for consolidating metadata for multiple files into a single file.

For more details on managing metadata for multiple files, see Data > Files > Save additional context.

Using the API

Creating a Model

You can create a new model by sending a POST request to https://app.valohai.com/api/v0/models/.

{
  "name": "<model-name>",
  "slug": "<model-uri-name>",
  "owner": <owner organization numeric ID>
}

Creating a Model Version

You can create a new model version from a list of datums by sending a POST request to https://app.valohai.com/api/v0/model-versions/.

{
  "model": "<model UUID>",
  "datums": ["<datum 1 UUID>", "<datum 2 UUID>", ...]
}

You can alternatively promote a dataset version into a model version with the following POST request structure:

{
  "model": "<model UUID>",
  "dataset_version": "<dataset version UUID>"
}